New Voice AI Skips the Typing Step, Answers Nearly 4x Faster
Your smart speaker could soon answer in half the time — even from a noisy kitchen.
When you ask a voice assistant a question, two things happen in sequence. First, software types out what you said — that's speech recognition. Then a second system searches a pile of documents for something relevant and reads the answer back. That two-step relay is slow, and it goes wrong in obvious ways: misheard words get carried into the search, so a question about "cedar trees" quietly becomes a search for "seeder trees." In a noisy kitchen, on a bad phone line, or with an accent the system doesn't expect, the whole chain falls apart.
The new method, called DUPAR, cuts out the middleman. Instead of converting your voice to text first, it compares the sound of your question directly against the written documents, using two search paths at once. A quick path checks a small memory of evidence found during your previous question, so follow-ups like "and how much does that cost?" get answered almost instantly. If that memory isn't confident enough, a slower path searches the full library and refills the memory for your next question.
The numbers are the interesting part. On a specialized knowledge base, the fast path was 3.75 times quicker than the standard "transcribe, then search" approach, while coming close to the accuracy of reading from clean text. In noisy conditions, the share of correct answers found in the top ten results rose from 0.771 to 0.875. Across different speaking styles, the top answer was correct 4.2 percentage points more often. It handles back-and-forth conversation — asking about a hotel, then about its parking — noticeably better than searching everything from scratch each time.
One honest caveat: this is a research paper, not a product you can use today. It was tested on one specialized set of documents with computer-generated speech, not on real phones in real homes. And it shines brightest when you ask follow-up questions rather than starting fresh. Still, the direction is clear. Removing the transcription step is one of the simplest ways to make voice assistants feel fast instead of sluggish.
- Voice assistants currently convert your speech to text before searching, which causes delays and misheard-word mistakes.
- The new method scored 3.75 times faster on the search step and answered more accurately in noisy conditions.
- It works best for follow-up questions, like asking about a hotel and then about its parking, because it remembers the previous answer.
Why It Matters
Faster, more accurate voice assistants save you time and frustration, especially in cars, kitchens, and noisy places.